AutoLTS: Automating Cycling Stress Assessment via Contrastive Learning and Spatial Post-processing

TL;DR

AutoLTS tackles the data-intense problem of cycling stress assessment by learning from street-view images through a two-step pipeline that first predicts road features with an ordinal-aware contrastive loss and then fuses these features with image embeddings to predict LTS. It introduces OrdCon to preserve ordinal relationships among LTS labels and a spatial post-processing module grounded in a causal model to enforce smoothness across connected road segments. On a Toronto dataset of 39,153 road segments, AutoLTS demonstrates strong LTS prediction performance, with substantial gains when street-view data are combined with partial road features and outperformance over baselines like MoCo and SupCon. The work enables scalable, image-based cycling-stress assessment with practical implications for urban planning and routing, while noting limitations such as domain shift and city-specific data needs for broader generalization.

Abstract

Cycling stress assessment, which quantifies cyclists' perceived stress imposed by the built environment and motor traffics, increasingly informs cycling infrastructure planning and cycling route recommendation. However, currently calculating cycling stress is slow and data-intensive, which hinders its broader application. In this paper, We propose a deep learning framework to support accurate, fast, and large-scale cycling stress assessments for urban road networks based on street-view images. Our framework features i) a contrastive learning approach that leverages the ordinal relationship among cycling stress labels, and ii) a post-processing technique that enforces spatial smoothness into our predictions. On a dataset of 39,153 road segments collected in Toronto, Canada, our results demonstrate the effectiveness of our deep learning framework and the value of using image data for cycling stress assessment in the absence of high-quality road geometry and motor traffic data.

Figures (6)

• Figure 1: Example images with the four LTS labels: LTS1 roads are safe for all cyclists including children, LTS2 roads are for most adults, LTS3 and LTS4 are for "enthused and confident" and "strong and fearless" cyclists, respectively.
• Figure 2: An overview of AutoLTS. The input image is encoded to an image embedding and is used to predict missing road features. The image encoder is trained using a contrastive learning approach (Section \ref{['subsec:contrastive']}). The predicted road features go through a post-processing module (Section \ref{['subsec:spatial_postprocessing']}) that enforces spatial smoothness into the predictions. Finally, a feedforward network predicts the the image's LTS label based on the image embedding, and the predicted and available road features.
• Figure 3: The contrastive learning framework and the learned image embeddings from different contrastive losses. MoCo indicates the self-supervised contrastive loss, SupCon indicates the supervised contrastive loss, and OrdCon indicates our contrastive loss. All the embeddings are projected to a two-dimensional space via T-SNE hinton2002stochastic. Each point corresponds to one street-view image and is color-coded according to the associated LTS label.
• Figure 4: A causal model for road feature predictions. The blue lines indicate real-world road segments, black arrows represent causal impacts.
• Figure 5: Illustration of the three spatial splits. York has a similar LTS distribution as the overall city-wide distribution. Etobicoke has the majority of the road segments being LTS2 and more roads being LTS4 compared to the city's average. Scarborough has an even higher LTS4 percentage.

Theorems & Definitions (2)

• Proposition 1
• proof